Can decision models play poker?

Part 1 · solved spots

The spot test

Each model saw the same turn and river situations and picked a move. We compared every pick with the move a poker solver (a perfect-play computer) would make.

Part 1b · five models, one decision

When all five agree

We asked five different AI models the same questions and looked at how many of them chose the same move — and whether the move most of them chose was any good.

Part 2 · full hands, every format

Real hands, cash to tournaments

Here the models play whole hands — every decision from the flop to the river, or every decision before the flop — in the formats people actually play. We count how many big blinds they give away per 100 hands compared with perfect play. Lower is better; 0 is perfect.

Postflop suites: flops solved from each format's preflop ranges, hands dealt and played to the end along the GTO line, every decision of both seats recorded. Preflop suites: heads-up blind-vs-blind / Spin HU trees from the preflop solver. Metric: bb/100 lost vs GTO = 100 × mean over hands of Σ EV loss ÷ 2 (the model plays both seats), 95% CI over hands. Hands shared by a suite and its dev/test splits are counted once.

Part 3 · the arena

Heads-up against real opponents

Perfect play is one yardstick; real opponents are another. Each model played hundreds of complete heads-up hands against two bots: a solid regular that plays tight and aggressive, and a calling station that never folds. Every deal is played twice with the seats swapped, so luck of the cards mostly cancels out. Higher is better; above 0 means the model won.

Duplicate heads-up matches, 100bb, every deal played twice with seats swapped. bb/100 ± 95% CI over duplicate pairs (pooled across matches of the same pairing); all-in EV bb/100 removes all-in luck. Matches with fewer than 100 hands and self-play are excluded; bot-vs-bot matches are references.

Part 4 · experiments

Pushing GPT-6 Luna

Can a model get better just by asking differently? We tried several ways of describing the table and asking the question. Each idea was tuned on one set of practice hands, then checked on fresh hands it had never seen — the only result that counts.

Variants of the request (state and question) and post-hoc players derived from stored answers, on the dev split (tuning) and the held-out test split (confirmation) of the cash hands suite. Δ = paired difference in bb/100 lost vs the default variant (v2) over the hands both answered, 95% CI. Negative is better. Hybrid players use non-model heuristics and are not pure model results.